Machine Learning Strategies for Augmented Health Informatics

这项研究计划聚焦于机器学习(ML)中与健康信息学应用密切相关的三个突出挑战：(1)机器学习有效运行的假设与现实世界应用程序中实际遇到的情况之间的不匹配；(2)缺乏针对特定情况选择适当的ML体系结构的系统指南；(3)根深蒂固的看法，即ML是一种黑箱方法，不利于人类理解。为了应对这些挑战，提出了三种相应的工程策略。对于第一个挑战，将研究数据平衡和生成性建模策略。特别是，ML方法通常取决于可用于成功部署的训练数据的数量和质量。这些假设在实际的健康应用中可能无法满足，因此严重影响了ML的适用性。然而，利用数据增强和生成性对抗网络等技术，在实际应用中开发ML初始化和训练解决方案以适应有限的数据集是可行的。对于第二个挑战，将研究允许基于资源可用性和遇到的环境进行系统的ML体系结构选择和适应的策略。为响应运行条件的变化，自适应ML策略应允许对系统参数进行有效调整和优化。因此，状态监测和反馈将被用来机会性地使ML系统适应在各种条件下的稳健性能。对于第三个挑战，将探索统称为可解释人工智能(XAI)的策略，以将人类设计因素纳入ML系统。这些策略寻求开发ML方法，这些方法不仅提供最终的输出决策，还提供相关的解释，以便人类能够理解和解释结果。在健康信息学应用中，ML系统通常必须在他们的协作努力中与人类交互，以使选定的治疗解决方案合理化。为此，XAI代表了一个很有希望的方向，以促进ML系统和人类用户之间的成功交互。总而言之，这些新的策略应该允许人类有效地利用ML进行决策，同时保持对所做决定的认知和控制，这些特征与新兴的增强智能(Ami)范式一致。特别是在健康信息学这一高度影响加拿大生活质量的应用领域，护理质量的进步是可以预期的：基于ML的疾病诊断和治疗不仅可以及时和准确，而且对人类来说也是相关和可解释的。因此，这些研究战略应该在提升加拿大作为健康和ML技术领导者的声誉方面发挥重要作用。

This research program focuses on three outstanding challenges in machine learning (ML), which are notably germane to health informatics applications: (1) a mismatch between assumptions for ML to operate effectively, and situations actually encountered in real-world applications; (2) a lack of systematic guidelines for selecting the appropriate ML architecture for a particular situation; (3) an entrenched perception that ML is a black-box approach, not conducive to human understanding. In order to tackle these challenges, three corresponding engineering strategies are proposed. For the first challenge, data balancing and generative modelling strategies will be investigated. In particular, ML methods are typically contingent on both the quantity and quality of available training data for successful deployment. These assumptions may not be met in practical health applications, thus severely compromising the applicability of ML. However, with techniques such as data augmentation and generative adversarial networks, it is feasible to develop ML initialization and training solutions for accommodating limited datasets in practical applications. For the second challenge, strategies allowing for systematic ML architecture selection and adaptation, based on resource availability and encountered environment, will be studied. In response to changes in operating conditions, the adaptive ML strategies should allow for effective tuning and optimization of system parameters. Accordingly, condition monitoring and feedback will be used to opportunistically adapt the ML systems for robust performance in various conditions. For the third challenge, strategies known collectively as explainable artificial intelligence (XAI) will be explored to incorporate human design factors into ML systems. These strategies seek to develop ML methods that provide not only the final output decisions but also the associated explanations, so that humans can understand and interpret the results. In health informatics applications, ML systems typically have to interact with humans in their collaborative efforts to rationalize a selected therapeutic solution. To this end, XAI represents a promising direction to facilitate the successful interaction between ML systems and human users. Together, these novel strategies should allow humans to efficiently utilize ML for decision making, while remaining cognizant and in control of the decisions made, characteristics which are consistent with the emerging paradigm of augmented intelligence (AmI). Specifically within health informatics, an application domain that highly impacts Canadian quality of life, advances in the quality of care can be expected: diagnosis and treatment of diseases based on ML can be not only timely and accurate, but also relevant and explainable to humans. Therefore, these research strategies should play a significant role in advancing Canada's reputation as a leader in both health and ML technologies.